← board

The shim slot is Pascal-only; a Python-shaped shim has nowhere honest to live

Implementation of decide-how-python-shaped-shims-should-be-shipped, decided by Rene 2026-08-17. Direction is settled; this is the work.

Why it is urgent rather than tidy

six gates 15 of 58 files in the NilPy corpus ladder — the single largest blocker, ahead of webencodings-as-a-unit (7), xml_dom (3) and warnings (3). Track N measured the ladder at HEAD and the walls are missing modules, not frontend bugs: three N fixes landed 2026-08-17 and the compile count did not move (4 pinned, 4 HEAD). Nobody can build six until this lands.

The defect

PyMimicShimExists (compiler/parser.inc:33206) probes exactly two paths:

lib/pcl/mimic_<name>.pas
lib/rtl/mimic_<name>.pas

six's entire content is Python aliases — text_type = str. Expressing "the str type object as a value" from Pascal is not something any shim does. A NilPy .py serves it perfectly and was verified byte-for-byte against CPython by Track B; it simply has nowhere to live under a name that says what it is.

The Pascal-only slot is an accident of when the lookup was written, not a design.

What to do

Probe mimic_<name>.py alongside mimic_<name>.pas in the same roots, and load it through the NilPy path. Keep the two spellings in one lookup rather than growing a second mechanism beside it.

Why NOT lib/**/six.py

It works, and it is refused on the owner's ruling that the tree must say what each shim is:

"resolution is a MAPPING, and the unit keeps OUR name… no file in the tree carries the upstream name." (Rene, 2026-07-26)

"not sneakily load other shims." (Rene, 2026-08-17)

A file named six.py satisfying import six while the build reports no shims is a dishonest tree, independent of any flag. Do not harden --no-shims as part of this — it means "unsupported: we provide no shims", nothing is guaranteed past the opt-out, and complicating the common path to defend an opt-in audit flag was explicitly rejected.

Gate

make compiler/pascal26 + tools/gate.sh quick. A .npy test importing a module served by a mimic_<name>.py, and the same import failing under --no-shims — which is the observable proving the shim is visible AS a shim rather than masquerading as a library module.

2026-08-17 (frank2, Track A) — RESOLVED. The cause section was stale; the real gap was narrower and deeper.

Premises re-derived before writing code, per the standing rule. Every symptom reproduced; the stated cause did not hold.

What was already true

A mimic_<name>.py in a library root already loaded and ran. Dropped lib/rtl/mimic_probe6.py, imported it from a .npy, and it printed the shim note, ran correctly, and was refused by --no-shims:

note: probe6 -> mimic_probe6 (shim, subset)
shim-answered / False

The reason: the shim mapping resolves by recursing through `ParseUsesUnit('mimic_'

The actual defect

PyMimicShimExists is asked exactly one thing — may a host C header win this name? — and it probed only .pas. So a .py shim was invisible to that question alone: fine for a plain name, but any name that also names a host header lost to the header, and the failure surfaced far away as undefined variable on the first attribute touched. Reproduced with a header-colliding name (/usr/include/search.h exists):

import search  ->  error: undefined variable (find)

…and a second, pre-existing bug in the same function

Fixing the extension alone did not fix it, which is what exposed the rest: PyMimicShimExists hard-coded CWD-relative lib/rtl/ and lib/pcl/ while the loader is ExeDir-anchored. Run the compiler from any directory but the repo root and the check answered "no shim" about a shim that then loaded fine. That bug was already there for .pas shims — import string from another CWD lost mimic_string.pas to /usr/include/string.h. Same class again for -Fu roots: the loader probes PasUnitDirs for mimic_<name>.py, the check did not.

Three spellings of one question, two of them wrong — the normalise-dont-special-case smell exactly. Rewritten as one root × all extensions helper (PyMimicShimAt) applied over every root the loader itself uses: -Fu roots, the ExeDir-anchored lib roots, then the CWD-relative fallback the stable binary needs.

Verified

Test: test/test_nilpy_shim_py.npy + fixture test/shims/mimic_search.py, enumerated in test-nilpy (both halves — load AND the --no-shims refusal). Self-contained via -Fu so it adds no permanent fixture to lib/** (Track B's file-lane).

Neither ruling was touched: no file in the tree carries an upstream name (the fixture is mimic_search.py), and --no-shims was not hardened — it refuses the .py shim through the existing mapping branch, unchanged.

Handoff

This unblocks the six half of feature-nilpy-six-and-warnings-shims (frank3, Track B): the slot now accepts lib/rtl/mimic_six.py. Shim CONTENT stays that ticket's — this one only made the slot able to hold it.

Note for whoever takes it: this is the third Track A ticket filed 2026-08-17 with no Track A worker staffed (with bug-a-a-python-module-s-identity-is-its-name-not- its-file and bug-a-synapse-tls-handshake-jumps-into-the-stack-inside-x509-verify- cert).

Log